boba

Agent-native bug bounty hunting framework.
Everything an LLM agent needs to go from target scope to submitted vulnerability report.

Quickstart · What It Does · Architecture · Agent Guide · Vision · Changelog

Bug bounty hunting is a closed-loop process: discover attack surface, interact with the target, test for vulnerabilities, chain findings, report. Human hunters do this with a browser, Burp Suite, and a dozen CLI tools held together by bash scripts and muscle memory.

LLM agents can reason and write — but they have no way to hold an authenticated session, replay a modified HTTP request, or diff two responses to confirm an IDOR. Boba gives them that. Every phase of the hunting lifecycle is exposed as a typed Python function with structured I/O: subprocess adapters normalize 9+ security tools into a uniform interface, a Playwright-based interaction layer replaces Burp Suite, and a SQLite-backed hunt context gives the agent persistent state across the entire engagement. Scope enforcement is framework-level — the agent cannot touch what it shouldn't.

Quickstart

pip install boba                 # CLI only
pip install 'boba[mcp]'          # + MCP server (boba-mcp)
pip install 'boba[oob]'          # + Interactsh OOB listeners
pip install -e '.[dev]'          # development install

playwright install chromium      # required for browser-based vuln tests
boba --help

External binaries (subfinder, httpx, nuclei, ffuf, ...) are not Python deps — install whichever tools you need from the External Tools table below.

# Create a hunt with scope, run recon
boba hunt create --name acme --scope scope.yaml
boba recon subdomains <id> --domain acme.com
boba recon hosts <id>
boba recon urls <id> --domain acme.com

# Test for vulnerabilities
boba test sqli <id> --url https://app.acme.com/search --param q
boba test xss <id> --url https://app.acme.com/search --param q
boba test auth <id> --endpoint https://app.acme.com/admin

# Analyze, chain, and report
boba analyze severity <id> --platform hackerone
boba analyze chain <id>
boba report draft <id> --finding-id 7
boba report format <id> --report-id 1 --platform hackerone

# Query what you've found
boba context stats <id> --format json

Every command supports --format json for agent consumption.

What It Does

Recon & Enumeration — Subdomain discovery, live host detection, port scanning, historical URL mining, tech fingerprinting, directory fuzzing, web crawling. 10 adapters wrapping subfinder, httpx, naabu, gau, waybackurls, whatweb, katana, ffuf, nuclei.

Web Interaction — Headless Playwright browser with real-time traffic interception. HTTP client with request(), replay(), compare(), fuzz() — four Burp Intruder attack types. Named auth sessions (bearer, basic, cookie, form login). OOB callback listeners via Interactsh for blind vuln detection.

Vulnerability Testing — 11 detection engines: IDOR (multi-account response diffing), SSRF (response content analysis + OOB callbacks), XSS (reflected + DOM-based via browser), SQLi (error/boolean/time-based with multi-DBMS payloads), auth bypass (JWT alg:none, claim escalation, endpoint ACL testing), CSRF (token validation + cross-origin), race conditions (concurrent request divergence), open redirect (external host detection), mass assignment (field persistence check), password reset (host header injection + rate limiting), AI/LLM prompt injection (exfiltration + override scoring). Nuclei adapter for template-based scanning with custom YAML support.

Analysis & Intelligence — Coverage tracking (tested vs. untested per endpoint per vuln class), finding deduplication (union-find with canonical selection), CVSS 3.1 scoring with platform-specific payout estimation (HackerOne, Bugcrowd), vulnerability chaining (8 chain rules — e.g., redirect + SSRF → P1), and attack path prioritization (rank untested endpoints by likelihood of vulnerability).

Reporting — Structured report drafting from findings and chains with auto-generated reproduction steps. Platform-specific formatting (HackerOne, Bugcrowd, generic markdown). PoC evidence packaging (HTTP request/response dumps, evidence.json, README).

Scope Enforcement — Default-deny model. Exclusions always win over inclusions. Domain wildcards, CIDR ranges, URL prefixes. Enforced at the adapter layer — the agent cannot touch out-of-scope targets regardless of what it tries to pass downstream.

Persistence — SQLite (WAL mode, FK constraints) backing 17 tables: subdomains, hosts, ports, URLs, technologies, directories, HTTP exchanges, sessions, findings, coverage, dedup groups, chains, OOB listeners, reports, tool runs. Upserts deduplicate via json_each() merges. Full hunt state survives across invocations — the agent never loses context.

Architecture

CLI / MCP Server / Agent SDK          <- exposure layer
        │
   Tool Functions                     <- composition (recon.py, vuln.py)
        │
   Adapter Layer                      <- normalize external tools
   ├── CLI Adapters (subprocess)         subfinder, httpx, nuclei, ffuf...
   └── Interaction Adapters              Playwright browser, HTTP client
        │
   Core                               <- scope engine, hunt context (SQLite), models

Every adapter follows a 6-phase lifecycle: find_binary → pre_filter → build_command → run_subprocess → parse_output → post_filter. This normalizes tools with wildly different I/O (JSON lines, JSON arrays, plain text, file-based output, stdin piping) into a uniform ToolResult. Tool functions compose one or more adapters with scope checking and context persistence. The CLI bridges async with asyncio.run().

External Tools

Boba wraps these — install whichever you need:

Development

pip install -e ".[dev]"

pytest                          # 840 tests
ruff check src/ tests/          # lint
ruff format --check src/ tests/ # format check

Python 3.11+. Fully async. Dataclasses (not Pydantic). Ruff for lint + format (line-length 100).

Roadmap

• V1 — Recon & enumeration (10 adapters, scope engine, persistence, CLI)
• V2 — Browser, HTTP interaction, vulnerability testing (5 vuln tools, Nuclei, sessions, OOB)
• V3 — Analysis, chaining, severity scoring, report generation (6 more vuln tools, CVSS 3.1, platform formatting)
• V4 — Recon breadth (parameter discovery, API mapping, secret scanning, GraphQL, ASN, cloud buckets)

License

Apache 2.0 — see LICENSE.